Accompanying the micro-fabrication and the complexity of the semiconductor manufacturing process, measurement and inspection using a scanning electron microscope (SEM) have become increasingly important for semiconductor manufacturing. To photograph high signa-to-noise ratio (SNR) images for precise measurement and precise inspection, conventional methods must reduce the noise by accumulating multiple low SNR images irradiated by an electron beam at the same point multiple times. However, such multiple irradiations increase sample damage and measurement and inspection turnaround time. To accelerate the turnaround time and improve performance of measurement and inspection using advanced image processing, we evaluated deep learning-based denoising algorithms that show significant denoising performance compared to the conventional method. Both supervised and self-supervised learning are mainstream deep learning denoising algorithms. The former requires advance preparation of high SNR images, unlike the latter. However, SEM can acquire higher SNR images by averaging more frames. In addition, the cost of acquisition of the training images can be ignored in the fabrication process when the number of training image sets is small. Therefore, in this study, we trained several supervised and self-supervised learning methods on a small dataset comprising hundreds of images and compared their results. The denoising performance for low SNR SEM images of semiconductor circuits was compared using peak signal-to-noise ratio (PSNR), similarity index measure (SSIM), image contrast, and critical distance (CD) values. The results showed that supervised learning achieved higher performance. In addition, we propose a new framework that conducts CD measurements using high SNR images during training and feeds the results back into the model optimization for supervised training. The results showed that the proposed method has potential to improve many metrics.
